app.py

import gradio as gr

import argparse
import datetime
import json
import os
import time

import gradio as gr
import requests
from PIL import Image

from q_align.model.builder import load_pretrained_model

from q_align.conversation import (default_conversation, conv_templates,
                                   SeparatorStyle)
from q_align.constants import LOGDIR
from q_align.utils import (build_logger, server_error_msg,
    violates_moderation, moderation_msg)

from q_align.evaluate.scorer import QAlignScorer, QAlignAestheticScorer, QAlignVideoScorer

import gradio as gr

def load_video(video_file):
    from decord import VideoReader
    vr = VideoReader(video_file)

    # Get video frame rate
    fps = vr.get_avg_fps()

    # Calculate frame indices for 1fps
    frame_indices = [int(fps * i) for i in range(int(len(vr) / fps))]
    frames = vr.get_batch(frame_indices).asnumpy()
    return [Image.fromarray(frames[i]) for i in range(int(len(vr) / fps))]


pretrained="q-future/one-align"
# 自动 device：HF 免费 Space 没 GPU → 走 CPU（会非常慢但能跑）
import torch
device = "cuda:0" if torch.cuda.is_available() else "cpu"
print(f"[OneScorer fork] device = {device}", flush=True)
tokenizer, model, image_processor, _ = load_pretrained_model(pretrained, None, "mplug_owl2", device=device)

iqa_scorer = QAlignScorer(tokenizer=tokenizer, model=model, image_processor=image_processor)
iaa_scorer = QAlignAestheticScorer(tokenizer=tokenizer, model=model, image_processor=image_processor)
vqa_scorer = QAlignVideoScorer(tokenizer=tokenizer, model=model, image_processor=image_processor)

scorers = {"Image Aesthetics (IAA)": iaa_scorer, "Image Quality (IQA)": iqa_scorer, "Video Quality (VQA)": vqa_scorer}

LEVELS = ["excellent (5)", "good (4)", "fair (3)", "poor (2)", "bad (1)"]
scores = [5,4,3,2,1]
def image_classifier(input_img, input_vid, scorer_type):
    if scorer_type is None:
        scorer_type = "Image Quality (IQA)"
    this_scorer = scorers[scorer_type]
    if input_vid is not None:
        input_ = load_video(input_vid)
    elif input_img is not None:
        input_ = [input_img]
    if "Video" in scorer_type:
        input_ = [input_]
    probs = this_scorer(input_).mean(0).tolist()
    prob_dict = {LEVEL: prob for LEVEL, prob in zip(LEVELS, probs)}
    score = sum([prob * score for score, prob in zip(scores, probs)])
    return prob_dict, score

title_markdown = ("""

<div style="width: 100%; text-align: center; margin:auto;">
      <img style="width: 100%" src="https://raw.githubusercontent.com/Q-Future/Q-Align/main/fig/onescorer.png">
</div> 

<h4 align="center"> If you like the OneScorer, please give us a star ✨ on <a href='https://github.com/Q-Future/Q-Align'>[GitHub]</a> for latest update.  </h4>

<h5 align="center">
    <div style="display:flex; gap: 0.25rem;" align="center">
        <a href='https://q-align.github.io'><img src='https://img.shields.io/badge/Homepage-green'></a>
        <a href='https://github.com/Q-Future/Q-Align'><img src='https://img.shields.io/badge/Github-Code-blue'></a>
        <a href="https://Q-Future.github.io/Q-Align/fig/Q_Align_v0_1_preview.pdf"><img src="https://img.shields.io/badge/Technical-Report-red"></a>
        <a href='https://github.com/Q-Future/Q-Align/stargazers'><img src='https://img.shields.io/github/stars/Q-Future/Q-Align.svg?style=social'></a>
    </div>
</h5>

""")


input_img = gr.Image(type='pil', label="Upload an Image")
input_vid = gr.Video(label="Upload a Video (will INGORE the image if a video is uploaded)",sources=["upload"])
radio = gr.Radio(["Image Aesthetics (IAA)", "Image Quality (IQA)", "Video Quality (VQA)"], label="Task", info="Which Scorer will you need?")

input_img = gr.Image(type='pil', label="Upload an Image")

labels = gr.Label(label="Probabilities of rating levels:")
number = gr.Number(label="Output score:", info="Range in [1,5]. Higher is better.", precision=4)
# Fork 修改：
#   1) 删 examples（原仓库 LFS 图未在 HF 上拉取，导致 Interface 启动时缓存 examples 失败）
#   2) cache_examples=False（保险）
#   3) device 改为 auto（兼容 CPU Space；推理会非常慢）
demo = gr.Interface(
    fn=image_classifier,
    inputs=[input_img, input_vid, radio],
    outputs=[labels, number],
    description=title_markdown,
    cache_examples=False,
    article=(
        "This is a fork of [teowu/OneScorer](https://huggingface.co/spaces/teowu/OneScorer) "
        "for personal experimentation. Removed broken example assets so the Space can boot. "
        "Backed by paper: *Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels* (ICML 2024). "
        "**Note**: this Space runs on free CPU; expect very slow (minutes) inference for a single image."
    ),
)
demo.launch(show_error=True, show_api=False)